Burstiness and perplexity without the maths

Perplexity is how predictable a text is, and burstiness is how unevenly it is written. Detector makers love both words, and you can get each of them without a formula.

Neither one tells you who wrote a text. They describe how a text feels to read, and people and machines can both produce either feeling.

Perplexity: how surprised is the reader

Imagine reading a sentence word by word and guessing the next word before you see it. If you guess right almost every time, the text has low perplexity: it is predictable. If the next word keeps surprising you, perplexity is high.

A language model does exactly this guessing, only with numbers. Text that a model writes tends to follow the words a model finds likely, so it tends to be predictable to another model. That is the whole idea behind many detectors.

But predictable does not mean machine. Instructions, legal forms, recipes and official letters are predictable on purpose. A careful writer who prefers plain, common words also lands on the predictable side. And a model can be asked to write in an odd, surprising way.

Burstiness: short bursts and long runs

Burstiness is about rhythm. A bursty text has short sentences next to long ones, a punchy paragraph after a heavy one. A flat text has sentences of about the same length, one after another.

People are usually bursty without trying. We get excited, we trail off, we stop short. Machine text often comes out flatter, with sentences and paragraphs of similar size. Flat does not prove a machine, though. A tidy report, a news item edited three times, a school essay with a strict template: all flat.

A tiny example

This is an example made up for illustration, not a measurement.

Flat: "The city built a new park. The park has many trees. The trees give shade in summer. Visitors enjoy the shade."

Bursty: "The city built a park. Trees, mostly - hundreds of them, planted so close that in July you walk from one patch of shade to the next and barely notice the heat."

Both are fine writing. The first is simply more even: four sentences of five or six words. The second mixes a four-word sentence with a long one. A detector looking at rhythm would see the difference, and you can see it too.

How AuthorText relates to these ideas

We do not compute perplexity with a language model. Our analysis uses simpler, visible signals, and one of them is the cousin of burstiness: how much sentence length varies across the text, and how much paragraph length varies. The lower the variation, the more points the text collects. These measures carry the most weight in the score.

Repetition is the other relative. Predictable text tends to reuse the same words, phrases and openings, so we count repeated patterns, stock phrases in four languages and connecting words packed into nearly every sentence. All of this is described in how to read the numbers in an AuthorText analysis.

The nice thing about visible signals is that you can look at them yourself. The evidence list shows the exact lines, so you can decide whether the flat rhythm is a flaw or just your genre.

Why each is weak evidence alone

First, genre. Polished human prose in news, business and science is flat and predictable, and our measurements show it gets flagged more often. Second, length. In our larger corpus test, texts of 80 to 299 words were close to random, and even at 300 words and more the result stayed modest. Third, editing: human text edited by a model, or paraphrased machine text, blurs both signals. Fourth, modern models can be prompted to vary their rhythm and wording.

This is why the honest framing is "share of text with signs of generation", not "percentage of AI". More on the failure modes in why modern AI detectors get it wrong.

What to do next

Try it on your own writing. Paste a text into the checker, open the evidence list and see where sentences or paragraphs are nearly the same length. If that is a fault of your writing, the fixes are in how to make your text sound more alive. If you are judging someone else's text, do not decide on rhythm or predictability: use the fair checklist and ask for drafts. Students who worry about a flag will find practical steps in AI detectors for students.